How to Read an AI Land Cover Map One Pixel at a Time

Use Dynamic World’s class probabilities and observation dates to investigate an apparent change in land cover.

A visitor examines transparent aerial-image layers above a miniature landscape.
A single map color hides the competition between classes at a field or shoreline edge.

A square on a satellite map turns from grass to crops between two dates. Has the land actually changed? An AI model can assign a land-cover label to each observation, but the highest-scoring class may win by a very small margin. The useful evidence is in the scores, the original image and the observations on either side of the apparent change.

What Dynamic World publishes

Dynamic World applies deep learning to Sentinel-2 imagery and provides land-cover predictions at 10-meter resolution. For each valid pixel it gives nine class probabilities and the label of the class ranked first. Its categories include trees, grass, crops, built surfaces and water. Each output is associated with an input Sentinel-2 observation; clouds and shadows are masked where identified. Dynamic World V1 data documentation Brown and colleagues, Dynamic World

A computer monitor displaying a map or aerial view of an area with various features and markings.
The leading class is a label for the observation at the scale of that pixel.

A 10-meter pixel represents a footprint roughly 100 square meters in area. An edge pixel can contain lawn, a path and overhanging branches. Its class probabilities describe the classifier’s support for the alternative labels. A tree probability of 0.6 does not mean a surveyor measured 60 square meters of tree canopy in that square.

A label can flip without a large change in evidence

Suppose the two leading scores on one date are 0.45 for grass and 0.43 for crops. Reversing those two scores on the following date changes the displayed color. That flip differs from a series of clear observations showing a large sustained shift. Check the margin between the two leading classes and inspect the underlying images for seasonal change, haze or a mixed field boundary.

An aerial view of a serene beach scene with buildings, a pool, and ocean waves.
Compare the class margins and the underlying images across dates.

The product documentation recommends considering the top class probability when selecting confident examples. A stricter rule removes some ambiguous observations from the analysis. Report how much usable area remains, since a map of only the easiest pixels may underrepresent precisely the field edges a reader cares about. Dynamic World V1 data documentation

What would establish a real change?

Look at several comparable dates and seek an appropriately dated field observation or higher-resolution image. Verify that the coordinates refer to the same ground. Describe the feature you can see: a roof appeared, a field flooded, or the surface became bare. Establishing ownership, zoning or the purpose of a building requires another record.

A large truck with an orange and black exterior is parked in a dimly lit warehouse or industrial setting.
A person reviewing the original scene can resolve some ambiguous classifications.

A Chesapeake Bay aerial map supplies a complementary view at a finer scale. What High-Resolution Aerial Images Tell Us About Land Cover

Try it with the evidence

Read a label change without the map colors

Source notes. Illustrative Dynamic World-style observations at a field edge. On date A, grass scores 0.45 and crops 0.43. On B, grass scores 0.43 and crops 0.45. The other seven classes sum to 0.12 on each date and no individual score exceeds 0.12. No field observation is supplied.

Prompt. Identify each winning class and its margin. Explain what the color change establishes and select two independent checks before recording a conversion to cropland.

Check your reading

A is grass and B is crops, each by 0.02. Only the order of the top two predictions is established. Check the underlying imagery and additional clear dates, then seek a dated field or higher-resolution reference for an actual conversion.

Illustrated field notes

Mapping a feature or an event
Aerial view of a city street at night with red and blue colors.
A building footprint is a different spatial object from a broad land-cover class.
A man is intently focused on analyzing data displayed on two computer monitors in a dimly lit office space.
Hazard likelihood needs more evidence than a land-cover category.
A collection of vibrant posters or prints featuring swirling patterns and shapes reminiscent of aurora borealis.
Preserve the date of the source when comparing a historic map with current imagery.
A man sits in front of multiple computer monitors, likely engaged in disaster response and management activities.
Damage mapping requires reference observations of the affected area.
Models, timing and decisions
A man is analyzing data on a large screen in an office setting.
A general-purpose image representation still needs a test for the local task.
A miniature cityscape on a table or desk, featuring roads, buildings, and trees.
A recent classification can still describe a cloudy or mixed scene imperfectly.
A person is holding an iPad with a photo of ancient ruins on its screen, possibly taken at an archaeological site or museum.
Keep planning boundaries separate from visible ground cover.